跪拜 Guibai
← All articles
AI Programming · AIGC

Tencent's Free Hy4 Flash Built a Game, Opened Its Own Browser, and Refused to Trust Its QA Pass

By 小虎AI生活 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent tooling just showed that a near-free model can handle multi-step autonomous work with real browser verification, but the hard-coding ceiling means teams should reserve it for exploratory and agentic tasks, not production engineering.

Summary

Tencent's Hy4 flash, free through November, got a single-sentence brief to build a browser game about a one-person company. The model didn't just start coding. It first set a design principle: a 12-slot × 30-day game must force a real trade-off between earning money and staying alive, or it's boring. It then wrote the game, probed the machine for browser automation, and played through the whole thing itself, clicking all 12 slots, all 5 actions, triggering the 3 random events, and reaching both endings.

The standout behavior: its own check program reported all passed, but the model still took screenshots to confirm layout and Chinese encoding. Before making the requested change (30 days/100k to 15 days/150k), it verified the new win condition is actually reachable. After editing, it re-ran a 15-item check and looked at the UI text. Tencent's own blind test claims satisfaction nearly matching Claude Haiku at a fraction of the cost, but the model scores only in the teens on hard engineering benchmarks.

Takeaways
— Hy4 flash is free until the end of November; Tencent's blind test says satisfaction nearly matched Claude Haiku at a fraction of the cost.
— Given a vague requirement ("set numbers yourself but reasonable"), the model first set a design principle rather than filling values blindly.
— It defined "done" as actually opening a browser and playing the game, probing for automation tools and writing a check program covering 12 slots, 5 actions, numbers, random events, and both endings.
— Even after its program reported all passed, it manually inspected screenshots for layout and Chinese encoding.
— It checked random events by looking at event log highlights and number changes, not just a "passed" string.
— Before modifying the win condition (30 days/100k → 15 days/150k), it verified the new condition is reachable, then re-ran a 15-item verification and confirmed the UI text changed.
— The author's verdict: multi-step, self-verifying, thinking-heavy tasks can be handed to Hy4 flash.
— Hy4 flash scored in the teens on hard engineering programming benchmarks, while its bigger sibling Hy4 scored 40+.
— The test is one developer's hands-on record; Tencent's blind test is self-reported.
Conclusions

The model's refusal to trust its own assertion program is a meaningful step: it grounds verification in the actual rendered UI, which catches the errors that unit tests miss (layout, encoding).

Reachability checking before a change is rare in generated code; treating "can the player still win?" as a first-class requirement prevents the model from producing a game that is technically correct but unwinnable.

The gap between strong agentic behavior and poor hard-engineering scores shows that free models can dominate bounded, environment-checkable tasks while still failing at open-ended production code; cost savings don't imply ceiling.

Since Tencent's blind test is self-reported and the hands-on test is a single case, the "satisfaction nearly matches Claude Haiku" claim needs independent verification before it enters procurement decisions.

Concepts & terms
Hy4 flash
Tencent's new lightweight model, free until the end of November, positioned for fast and cheap tasks like document processing, web browsing, and simple agentic workflows; it scores low on hard engineering benchmarks.
WorkBuddy
A tool that gives AI agents a sandboxed environment with file access and browser automation, letting models write, run, and visually verify their own output.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗